As mentioned in the comments, your biological question is going to be vitally important in determining the best way to tackle this issue. However, there are some things that you should generally keep in mind.
Most de novo assemblers will perform kmer normalization. This means that the merged RNA-Seq data will be normalized across all libraries. This is almost certainly not what you want because you'll be randomly selecting reads from all libraries without regard to their treatments. So, if you're going to allow for kmer normalization, you should normalize prior to merging the libraries so that you have equal representation of the reads by library. If you don't normalize, you're going to have a very fragmented assembly, and it will take a long time to generate.
De novo assemblers using short reads assume all identified exons with junction support may be used to generate splice variants with equal probability. I'm not entirely sure how to best phrase that, but here's an example. Suppose in a cold-treated library that Gene A generates a transcript using exons 1-2-3, but a heat-treated library generates transcripts using 3-4-5 and 2-3-5. It may be that merging the libraries results in transcripts generated containing various permutations of those exons (e.g. 1-2-3-5, 2-3-4, etc.) that aren't generated in any of the biological settings but can no longer be distinguished. This will lead to false positive transcripts which will detrimentally affect expression quantification.
Merging assemblies is difficult. So, to be true to each treatment, it makes more sense to generate individual assemblies (which will also result in library-specific kmer normalization). StringTie essentially does this using reference genomes. You assemble using each library (all bio-reps for a treatment), and then merge after the fact and re-quantify. This is the best approach. However, it can be difficult and computationally taxing. Some programs like CD-HIT (specifically cd-hit-est) are typically used in these cases. Keep in mind that CD-HIT will drop sequences with high similarity, not merge them. This may lead to loss of some uniquely assembled sequence. Other approaches involve using programs like CAP3 (a DNA assembler) to treat the transcripts like longer reads for assembly. Depending on the parameter choice, this might work, but you will risk generating chimeric transcripts.
How you move forward with the assembly process will again depend on your biological questions. I would tend to promote that you tackle the questions one at a time rather than using a single transcriptome to address them all, but there may be circumstances in which merging assemblies makes more sense.
In your specific case, I would perform multiple assemblies using a variety of the approaches described above and then annotate the assemblies (ex. using Trinotate) to see how their quality metrics compare. I would discourage generating a single transcriptome and then re-assembling individual transcriptomes from that assembly -- the single transcriptome may have misassemblies that would then propagate to the individual assemblies, and this process would generally be messy/difficult.
